-
Notifications
You must be signed in to change notification settings - Fork 29
/
Copy pathtrain.py
208 lines (173 loc) · 8.42 KB
/
train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
import torch.optim as optim
import torch
from utils.train_utils import *
import logging
import math
import importlib
import datetime
import random
import munch
import yaml
import os
import sys
import argparse
from dataset import ShapeNetH5
def train():
logging.info(str(args))
metrics = ['cd_p', 'cd_t', 'emd', 'f1']
best_epoch_losses = {m: (0, 0) if m == 'f1' else (0, math.inf) for m in metrics}
train_loss_meter = AverageValueMeter()
val_loss_meters = {m: AverageValueMeter() for m in metrics}
dataset = ShapeNetH5(train=True, npoints=args.num_points)
dataset_test = ShapeNetH5(train=False, npoints=args.num_points)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=args.batch_size, shuffle=True, num_workers=int(args.workers))
dataloader_test = torch.utils.data.DataLoader(dataset_test, batch_size=args.batch_size, shuffle=False, num_workers=int(args.workers))
logging.info('Length of train dataset:%d', len(dataset))
logging.info('Length of test dataset:%d', len(dataset_test))
if not args.manual_seed:
seed = random.randint(1, 10000)
else:
seed = int(args.manual_seed)
logging.info('Random Seed: %d' % seed)
random.seed(seed)
torch.manual_seed(seed)
model_module = importlib.import_module('.%s' % args.model_name, 'models')
net = torch.nn.DataParallel(model_module.Model(args))
net.cuda()
if hasattr(model_module, 'weights_init'):
net.module.apply(model_module.weights_init)
cascade_gan = (args.model_name == 'cascade')
net_d = None
if cascade_gan:
net_d = torch.nn.DataParallel(model_module.Discriminator(args))
net_d.cuda()
net_d.module.apply(model_module.weights_init)
lr = args.lr
if cascade_gan:
lr_d = lr / 2
if args.lr_decay:
if args.lr_decay_interval and args.lr_step_decay_epochs:
raise ValueError('lr_decay_interval and lr_step_decay_epochs are mutually exclusive!')
if args.lr_step_decay_epochs:
decay_epoch_list = [int(ep.strip()) for ep in args.lr_step_decay_epochs.split(',')]
decay_rate_list = [float(rt.strip()) for rt in args.lr_step_decay_rates.split(',')]
optimizer = getattr(optim, args.optimizer)
if args.optimizer == 'Adagrad':
optimizer = optimizer(net.module.parameters(), lr=lr, initial_accumulator_value=args.initial_accum_val)
else:
betas = args.betas.split(',')
betas = (float(betas[0].strip()), float(betas[1].strip()))
optimizer = optimizer(net.module.parameters(), lr=lr, weight_decay=args.weight_decay, betas=betas)
if cascade_gan:
optimizer_d = optim.Adam(net_d.parameters(), lr=lr_d, weight_decay=0.00001, betas=(0.5, 0.999))
alpha = None
if args.varying_constant:
varying_constant_epochs = [int(ep.strip()) for ep in args.varying_constant_epochs.split(',')]
varying_constant = [float(c.strip()) for c in args.varying_constant.split(',')]
assert len(varying_constant) == len(varying_constant_epochs) + 1
if args.load_model:
ckpt = torch.load(args.load_model)
net.module.load_state_dict(ckpt['net_state_dict'])
if cascade_gan:
net_d.module.load_state_dict(ckpt['D_state_dict'])
logging.info("%s's previous weights loaded." % args.model_name)
for epoch in range(args.start_epoch, args.nepoch):
train_loss_meter.reset()
net.module.train()
if cascade_gan:
net_d.module.train()
if args.varying_constant:
for ind, ep in enumerate(varying_constant_epochs):
if epoch < ep:
alpha = varying_constant[ind]
break
elif ind == len(varying_constant_epochs)-1 and epoch >= ep:
alpha = varying_constant[ind+1]
break
if args.lr_decay:
if args.lr_decay_interval:
if epoch > 0 and epoch % args.lr_decay_interval == 0:
lr = lr * args.lr_decay_rate
elif args.lr_step_decay_epochs:
if epoch in decay_epoch_list:
lr = lr * decay_rate_list[decay_epoch_list.index(epoch)]
if args.lr_clip:
lr = max(lr, args.lr_clip)
for param_group in optimizer.param_groups:
param_group['lr'] = lr
for i, data in enumerate(dataloader, 0):
optimizer.zero_grad()
if cascade_gan:
optimizer_d.zero_grad()
_, inputs, gt = data
# mean_feature = None
inputs = inputs.float().cuda()
gt = gt.float().cuda()
inputs = inputs.transpose(2, 1).contiguous()
# out2, loss2, net_loss = net(inputs, gt, mean_feature=mean_feature, alpha=alpha)
out2, loss2, net_loss = net(inputs, gt, alpha=alpha)
if cascade_gan:
d_fake = generator_step(net_d, out2, net_loss, optimizer)
discriminator_step(net_d, gt, d_fake, optimizer_d)
else:
train_loss_meter.update(net_loss.mean().item())
net_loss.backward(torch.squeeze(torch.ones(torch.cuda.device_count())).cuda())
optimizer.step()
if i % args.step_interval_to_print == 0:
logging.info(exp_name + ' train [%d: %d/%d] loss_type: %s, fine_loss: %f total_loss: %f lr: %f' %
(epoch, i, len(dataset) / args.batch_size, args.loss, loss2.mean().item(), net_loss.mean().item(), lr) + ' alpha: ' + str(alpha))
if epoch % args.epoch_interval_to_save == 0:
save_model('%s/network.pth' % log_dir, net, net_d=net_d)
logging.info("Saving net...")
if epoch % args.epoch_interval_to_val == 0 or epoch == args.nepoch - 1:
val(net, epoch, val_loss_meters, dataloader_test, best_epoch_losses)
def val(net, curr_epoch_num, val_loss_meters, dataloader_test, best_epoch_losses):
logging.info('Testing...')
for v in val_loss_meters.values():
v.reset()
net.module.eval()
with torch.no_grad():
for i, data in enumerate(dataloader_test):
label, inputs, gt = data
# mean_feature = None
inputs = inputs.float().cuda()
gt = gt.float().cuda()
inputs = inputs.transpose(2, 1).contiguous()
# result_dict = net(inputs, gt, is_training=False, mean_feature=mean_feature)
result_dict = net(inputs, gt, is_training=False)
for k, v in val_loss_meters.items():
v.update(result_dict[k].mean().item())
fmt = 'best_%s: %f [epoch %d]; '
best_log = ''
for loss_type, (curr_best_epoch, curr_best_loss) in best_epoch_losses.items():
if (val_loss_meters[loss_type].avg < curr_best_loss and loss_type != 'f1') or \
(val_loss_meters[loss_type].avg > curr_best_loss and loss_type == 'f1'):
best_epoch_losses[loss_type] = (curr_epoch_num, val_loss_meters[loss_type].avg)
save_model('%s/best_%s_network.pth' % (log_dir, loss_type), net)
logging.info('Best %s net saved!' % loss_type)
best_log += fmt % (loss_type, best_epoch_losses[loss_type][1], best_epoch_losses[loss_type][0])
else:
best_log += fmt % (loss_type, curr_best_loss, curr_best_epoch)
curr_log = ''
for loss_type, meter in val_loss_meters.items():
curr_log += 'curr_%s: %f; ' % (loss_type, meter.avg)
logging.info(curr_log)
logging.info(best_log)
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='Train config file')
parser.add_argument('-c', '--config', help='path to config file', required=True)
arg = parser.parse_args()
config_path = arg.config
args = munch.munchify(yaml.safe_load(open(config_path)))
time = datetime.datetime.now().isoformat()[:19]
if args.load_model:
exp_name = os.path.basename(os.path.dirname(args.load_model))
log_dir = os.path.dirname(args.load_model)
else:
exp_name = args.model_name + '_' + args.loss + '_' + args.flag + '_' + time
log_dir = os.path.join(args.work_dir, exp_name)
if not os.path.exists(log_dir):
os.makedirs(log_dir)
logging.basicConfig(level=logging.INFO, handlers=[logging.FileHandler(os.path.join(log_dir, 'train.log')),
logging.StreamHandler(sys.stdout)])
train()